The Performance of School Assignment Mechanisms in Practice

DISCUSSION PAPER SERIES IZA DP No. 9118 The Performance of School Assignment Mechanisms in Practice Monique de Haan Pieter A. Gautier Hessel Oosterbeek Bas van der Klaauw June 2015 Forschungsinstitut zur Zukunft der Arbeit Institute for the Study of Labor

The Performance of School Assignment Mechanisms in Practice Monique de Haan University of Oslo Pieter A. Gautier VU University Amsterdam and IZA Hessel Oosterbeek University of Amsterdam Bas van der Klaauw VU University Amsterdam and IZA Discussion Paper No. 9118 June 2015 IZA P.O. Box 7240 53072 Bonn Germany Phone: +49-228-3894-0 Fax: +49-228-3894-180 E-mail: iza@iza.org Any opinions expressed here are those of the author(s) and not those of IZA. Research published in this series may include views on policy, but the institute itself takes no institutional policy positions. The IZA research network is committed to the IZA Guiding Principles of Research Integrity. The Institute for the Study of Labor (IZA) in Bonn is a local and virtual international research center and a place of communication between science, politics and business. IZA is an independent nonprofit organization supported by Deutsche Post Foundation. The center is associated with the University of Bonn and offers a stimulating research environment through its international network, workshops and conferences, data service, project support, research visits and doctoral program. IZA engages in (i) original and internationally competitive research in all fields of labor economics, (ii) development of policy concepts, and (iii) dissemination of research results and concepts to the interested public. IZA Discussion Papers often represent preliminary work and are circulated to encourage discussion. Citation of such a paper should account for its provisional character. A revised version may be available directly from the author.

IZA Discussion Paper No. 9118 June 2015 ABSTRACT The Performance of School Assignment Mechanisms in Practice * Theory points to a potential trade-off between two main school assignment mechanisms; Boston and Deferred Acceptance (DA). While DA is strategy-proof and gives a stable matching, Boston might outperform DA in terms of ex-ante efficiency. We quantify the (dis)advantages of the mechanisms by using information about actual choices under Boston complemented with survey data eliciting students school preferences. We find that under Boston around 8% of the students apply to another school than their most-preferred school. We compare allocations resulting from Boston with DA with single tie-breaking (one central lottery; DA-STB) and multiple tie-breaking (separate lottery per school; DA-MTB). DA-STB places more students in their top-n schools, for any n, than Boston and results in higher average welfare. We find a trade-off between DA-STB and DA-MTB. DA-STB places more students in their single most-preferred school than DA-MTB, but fewer in their top-n, for n 2. Finally, students from disadvantaged backgrounds benefit most from a switch from Boston to any of the DA mechanisms. JEL Classification: C83, D47, I20 Keywords: school choice, Boston mechanism, deferred acceptance mechanism, strategic behavior, ex-ante efficiency, ex-post efficiency Corresponding author: Bas van der Klaauw Department of Economics VU University Amsterdam De Boelelaan 1105 1081 HV Amsterdam The Netherlands E-mail: b.vander.klaauw@vu.nl * We benefited from valuable comments from Yinghua He, Fuhito Kojima and Alvin Roth, and from seminar participants at various places. We gratefully acknowledge the cooperation of secondary schools in Amsterdam.

1 Introduction In a system of free school choice, students can choose in which school to enroll subject to schools capacity constraints. Some schools may be oversubscribed while others have empty seats. Many cities around the world, therefore, operate a centralized school assignment mechanism. The most well-known mechanisms are the Boston mechanism and the Deferred Acceptance (DA) mechanism. An extensive literature has analyzed the theoretical properties and shows that these two main mechanisms have important advantages and disadvantages. (e.g. Gale and Shapley, 1962; Roth, 2008; Erdil and Ergin, 2008; Abdulkadiroǧlu et al., 2009; Che and Kojima, 2010; Kojima and Manea, 2010; Abdulkadiroǧlu et al., 2011). 1 Most prominently, under the Boston mechanism students do not necessarily reveal their true preferences (the mechanism is not strategy proof). If students have a low acceptance probability at their most-preferred school, for example because the school is popular or they do not have priority at this school, they may rank another school first. The lack of strategy-proofness as property of the Boston mechanism is particularly problematic if the ability to strategize differs by social background. In that case students for disadvantaged backgrounds are more likely to end up at an undesired school than other students. In contrast, the DA mechanism is strategy-proof. It is optimal for students to reveal their true preferences, because a rejection at the most-preferred school does not reduce the acceptance probability at the next school on a student s preference list. Furthermore, DA results in a stable matching, because no student looses a seat at a (preferred) school to a student who has lower priority at that school. The Boston mechanism does not have this property of a stable matching, but this does not imply that the DA is superior to the Boston mechanism. Abdulkadiroǧlu et al. (2011) show that the Boston mechanism can dominate the DA mechanism on the basis of ex-ante efficiency, because the Boston mechanism gives students some scope to express the intensity of their preferences by ranking their most-preferred school first even when the probability of being rejected is high. The DA mechanism does not give an opportunity to signal the intensity of preferences. 2 As a result, the 1 A third mechanism discussed in the literature is the Top Trading Cycles mechanism (TTC). We do not analyze this mechanism here because it is considered undesirable when schools use priority rules (as is the case in our setting). TTC gives students who have priority at a school, for example because an older sibling is enrolled in that school, a higher probability to be placed in any school than students who do not have priority at any school. 2 This is addressed in the Choice-Augmenented DA mechanism where students can signal their cardinal preferences by sending an additional message indicating their "target school" which is used to break ties at schools. The choice of a target school is, however, a strategic choice (see Abdulkadiroǧlu et al. (2014)). 2

Boston mechanism may result in higher average welfare than the DA mechanism. It is important to quantify the trade off between Boston and DA, especially for policy makers who have to choose which school assignment mechanism to use. To do so, we have collected a unique data set in Amsterdam that contains information on actual school choices made under (a variant) of the Boston mechanism and information on the ordinal and cardinal preferences of students over schools. With this data set we can: i) obtain an estimate of the degree of strategic behavior under the Boston system, ii) investigate whether strategic behavior depends on characteristics of students and/or the intensity of preferences, and iii) compare the performance of the Boston and DA mechanisms in terms of ex-ante and ex-post efficiency. Ex-post efficiency is defined in terms of counting how often a student has a more favorable outcome under any of the DA mechanisms than under Boston. Ex-ante efficiency is based on the expected welfare of a student under the different mechanisms. So ex-post efficiency considers an ordinal measure of the student s preference, while ex-ante efficiency is based on cardinal measure for these preferences. Assessing whether students behave strategically and the properties of the resulting allocation, is not easy. Prior research has relied on laboratory experiments or on estimation of structural models. Chen and Sönmez (2006) are the first to examine school assignment mechanisms in the lab. Participants were given valuations for hypothetical schools and would earn an amount equal to the value of the school where they get a place. Participants in different sessions were confronted with different assignment mechanisms. Truth telling was more often observed under DA and Top- Trading Cycles (TTC) than under Boston. In contrast to theoretical predictions, DA Pareto dominates TTC, which in turn dominates Boston (see also Calsamiglia et al., 2011). He (2012); Calsamiglia et al. (2014); Agarwal and Somaini (2014) use structural models of school choice in cases where the Boston mechanism applies, to uncover students true preferences. They use the estimated preferences to simulate counterfactual assignment mechanisms. He (2012) uses data of almost one thousand students applying to four middle schools in a neighborhood in Beijing that uses a version of the Boston system. In his model agents play Bayesian-Nash equilibrium strategies but are allowed to have heterogenous beliefs. He finds that a change from Boston to DA has more losers than winners and that average welfare is lower under DA than under Boston. Calsamiglia et al. (2014) use data from students applying to elementary schools in Barcelona where students can submit a list of up to ten schools and may have priority based on distance and older siblings. Parents are sophisticated or naive, and are assumed to play equilibrium strategies, where sophisticated 3

parents have correct beliefs about admission probabilities. They find that over 90 percent of the parents are sophisticated and that a change from Boston to DA has more losers than winners, while a change from Boston to TTC has more winners than losers. Finally, Agarwal and Somaini (2014) use data from elementary school students in Cambridge (MA). They analyze the benchmark case where all parents are sophisticated and play a Bayesian-Nash equilibrium. This study finds that under Boston 84 percent of the students are assigned to their first-ranked school, while only 75 percent of the students are assigned to their most-preferred school. Furthermore, average welfare is lower under DA than under Boston. Like the structural papers we use administrative data from a city that assigns students to schools on the basis of (a variant of) the Boston mechanism. However, we do not make assumptions on choice behavior to uncover preferences. Instead, we combine the administrative data with data from questionnaires through which we elicited ordinal and cardinal preferences. Our approach assumes that stated preferences are an accurate measure of actual preferences. We consider this approach as complementary to structural analyses. An advantage of the stated preference approach is that it is more transparent. This turned out to be important when communicating our findings regarding counterfactual assignment mechanisms to policy makers. In 2013, we complemented the existing application procedure of students choosing a secondary school in Amsterdam with a questionnaire. In this questionnaire students were asked to submit a preference list of up to ten schools and to award preference points to the schools they listed. An accompanying letter emphasized that the responses to the questionnaire would not influence the current application procedure. If some students did not believe this promise, our findings are underestimating the true difference between Boston and DA. We compare Boston with two different versions of the (student-proposing) DA mechanism: DA where one centralized lottery breaks ties among students in the same priority group (single tie-breaking; DA-STB) and DA where each school runs its own lottery to break ties among students in the same priority group (multiple tie-breaking; DA-MTB). The main findings of our analyses are as follows. Around eight percent of the students disguise their true preferences under the Boston mechanism and apply to another school than their most-preferred school. This does not differ between boys and girls, nor between students from disadvantaged and non-disadvantaged neighborhoods. Applying to another school than the most-preferred school is, however, more likely for students who report a small difference in preference points between their most-preferred school and their second most-preferred school. This concurs 4

with the hypothesis that the likelihood of strategic behavior depends on the intensity of preferences. In terms of ex-post efficiency, DA-STB dominates Boston, that is, if we evaluate the mechanisms in terms of how many students are assigned to one of their mostpreferred n schools, DA-STB does better than Boston for any value of n. Both Boston and DA-STB assign more students to their single most-preferred school compared to DA-MTB, but DA-MTB assigns more students to one of their two (or higher) most-preferred schools than DA-STB and (in that order) Boston. To estimate the degree of ex-post inefficiency of the three mechanisms we simulate the fraction of students that would like to switch places without harming other students. The fraction of switchers is highest under DA-MTB (6%) followed by Boston (4%), while under DA-STB almost no switches are possible without harming other students. In terms of ex-ante efficiency, there is not one mechanism that dominates the others. In terms of expected preference points, a majority of the students is better off under Boston than under DA-STB and DA-MTB, while the average number of preference points is higher under DA-STB than under DA-MTB and Boston. We further find that students from disadvantaged backgrounds gain more (lose less) from a switch from Boston to DA-STB or DA-MTB. The remainder of the paper is organized as follows. Section 2 provides a brief review of the properties of the assignment mechanisms that we focus on in this paper. Section 3 describes the main features of the secondary education system in the Netherlands and describes the Amsterdam version of the Boston mechanism. Section 4 describes the data that we collected for this study. Section 5 presents and discusses the results. Section 6 summarizes and concludes. 2 School assignment mechanisms This section provides a brief review of the Boston mechanism (Section 2.1) and of the DA mechanism (Section 2.2). 2.1 The Boston mechanism The Boston mechanism works as follows. Students submit an ordered list with their top n schools. Schools rank students on the basis of priority, a lottery makes the priority order strict. In the first round students are assigned to the school which they ranked first. Schools with more students than places reject the lowest ranked students. In the second round students who are rejected in the first round are 5

assigned to the school they ranked second if this school has remaining places. Again, if more students apply than there are remaining places, the lowest-ranked students of those who apply in the second round, are rejected. This process continues until no more students are rejected. The Boston mechanism has three properties that are considered undesirable. It is: i) not truth-telling, ii) not stable, and iii) not ex-post Pareto efficient. Not truthtelling means that students have an incentive to apply for a less-preferred school if they perceive that the probability of getting a place at their most-preferred school is low. Not stable (or justified envy) means that a student might be assigned to school s while she prefers s and has a higher priority at s than another student assigned to s. Not Pareto efficient means that in the resulting allocation, students might be better off by switching schools without harming other students. The Boston mechanism maximizes the number of students who are placed in the school they ranked first. Given that students may behave strategically, this does not maximize the number of students assigned to their most-preferred school. A potential advantage of the Boston mechanism pointed out by Abdulkadiroǧlu et al. (2011) is that it gives students some scope to express the intensity of their preferences. If two students prefer school s to school s but student i has a strong preference for s while student i is almost indifferent, student i is more likely to get a seat at s because student i has an incentive to "strategically" apply for school s. 2.2 The DA mechanism The (student-proposing) DA mechanism works as follows. 3 Students submit an ordered list with their top n schools. 4 Schools rank students on the basis of priority, a lottery makes the priority order strict. The lottery can be centralized (single tie-breaking; STB) or there can be separate lotteries at each school (multiple tiebreaking; MTB). In the first round students are tentatively assigned to the school they have ranked first. Schools with more students than places reject the lowestranked students. Students who are rejected in the first round are tentatively assigned to the second school on their choice list. Each school then considers students assigned in round 2 and the students it has been holding. Schools with more students than places reject the lowest-ranked students. Hence an oversubscribed school can in the second round reject students it was holding after the first round. This process continues until no students are rejected anymore. 3 The algorithm is due to Gale and Shapley (1962), and first proposed to assign students to schools by Abdulkadiroǧlu and Sönmez (2003). 4 Where n is large enough to avoid that students strategize regarding the choice of which schools to include in the list (cf. Calsamiglia et al., 2010). 6

DA is strategy-proof (or truth-telling). Since students are tentatively assigned in each step, there is no cost to listing true preferences. A student who is rejected at her favorite school in the first round can still be placed at a school where she has priority in a later round. DA results in a stable allocation. No student loses a seat to a lower priority student and is assigned to a less-preferred school. The allocation is, however, not ex-post Pareto efficient. The intensity of preferences plays no role in the DA mechanism. Students list their true ordinal preferences and if two students tie in priority at a school, a lottery and not the cardinal preferences determines who is accepted. 3 Context 3.1 Secondary education in the Netherlands Our data are from students who are finishing primary school and are choosing a secondary school. At this stage, students are 11 or 12 years old. The Netherlands has a tracked secondary education system. The lowest track (pre-vocational secondary education) lasts four years and gives access to subsequent vocational education programs. The intermediate track (general secondary education) takes five years and gives access to professional colleges. The highest track (pre-university education) takes six years and gives access to university education. Which track a student takes is determined at the end of primary education, and depends on the result of a nationwide exit test and on the advice of the primary school teacher. Students can freely choose among the schools that offer the track at their level. Virtually all schools are publicly funded and there are no substantial tuition fees. All schools prepare their students for nationwide exams at the end of secondary education. The Education Inspectorate assesses the quality of schools and publishes its findings on the Internet. Schools that receive the lowest quality score ( very weak ) for three years in a row are closed (if publicly-run) or lose their public funding (if privately-run). 3.2 School choice in Amsterdam Amsterdam is the capital of the Netherlands and is with 750,000 inhabitants its largest city. Each year, around 8000 students transfer from primary education to secondary education. In the city of Amsterdam, there are around 70 secondary schools. Since 2005, the secondary schools in Amsterdam run a centralized application and admission system. In the first round students can only apply to one school that 7

offers their advised track. Schools are allowed to use a limited number of priority rules, i.e. they can grant priority to siblings of current students, to children of staff members, and to students from primary schools with the same pedagogical approach (for example, Montessori or Dalton schools). 5 The priority rules need to be announced before the application date, so they are known to potential applicants. When a school is oversubscribed, students are admitted on the basis of priority. A school-based lottery determines the ordering of students who are not in a priority group. Schools may be oversubscribed for some school tracks but not for others. Lotteries are conducted for each school track separately. After losing a lottery, students can only apply to one of the schools that still have places available after the first round. When there are too many applicants for the places that a school still has available in the second round, a lottery is used to determine which of the secondround applicants are admitted. After losing a lottery in the second round, students should again apply to a school that still has seats available after this second round. In the third round, schools consider applications in chronological order. Finally, students who are not placed after the third round, are assigned to a school by a committee. For each track, the total capacity of all schools exceeds the total number of students assigned to this track in Amsterdam. The system used in Amsterdam is not identical but very similar to the Boston mechanism described in Section 2.1. Under the Boston mechanism, students who are rejected in round k skip round k + 1 if the k + 1 th school on their ordered list has no seats available at that stage. Because the Amsterdam mechanism only asks students to choose one school at the time, rejected students never skip a round. In spite of this difference, the Amsterdam mechanism has similar properties as the Boston mechanism: it is not strategy-proof, the allocation is neither stable nor Pareto efficient, and students have some scope to express the intensity of their preferences. Our study pertains to the cohort that enrolled in secondary school in September of 2013. These students received the advice from their primary school teacher around December 2012, they participated in the nationwide exit test in February 2013 and received the result of that test in March 2013. The secondary schools organized open days during which prospective new students can visit the schools and gather information between January 7 2013 and March 1 2013. Each student received exactly one official application form and had to submit that at the school where (s)he applied between March 4 2013 and March 15 2013. Since truth-telling is not a dominant strategy under the Boston mechanism, one 5 Priority on the basis of distance or walking zone is not allowed. 8

may benefit from knowing at which schools other students apply. In that regard it is important to note that some secondary schools in Amsterdam maintained a website where they report the number of applications they have received so far. Moreover, students who have already submitted their application form to a school can within the application window withdraw it from that school and submit it at another school. 6 These features of the procedure in Amsterdam may help students to coordinate their choices and are, therefore, likely to lead to a better allocation of students to schools. 4 Data Our data describe students who applied for secondary school in Amsterdam in 2013. The data come from two sources. The first source is the register of the centralized application and admission system of the city of Amsterdam, which contains all students. For each student we observe the primary school they attended, the track advised by the primary-school teacher, the score on the final exit test from primary school, the schools to which the student applied in the first and possibly later rounds, whether the student has priority at the school at which (s)he applied, and whether the student was admitted to the school or lost the lottery. The second source of data is a questionnaire that we administered alongside with the application procedure. All students who were in the final grade of primary school received an application form for secondary education, which they handed in at the school at which they applied. Together with this official application form, students received a questionnaire and an explanatory letter from us. In the letter we explain that the questionnaire is for research purposes to find out whether the current assignment procedure can be improved, and that the responses to the questionnaire do not influence the outcomes of the current procedure. It also emphasizes that data will be treated confidential and that we will not report data on individual students. Appendix A contains a translation of the explanatory letter from Dutch. The questionnaire is brief and asks to which school the student applied, what the reasons are for applying to this school, whether the student has priority at the school where (s)he applied, whether it is felt that there is a risk of losing the lottery and if so whether this chance is high or low, whether the student would apply to the same school if no single school would conduct a lottery, and how many schools were visited before choosing a school. In addition to these questions, the questionnaire asks students to make a preference list of up to ten schools, and to award points 6 This requires physically going to these schools. 9

to each school on the list. Students were asked to hand in the questionnaire (in a return envelop addressed to us) at the school where they applied in the first round. Students who did not bring a questionnaire were supposed to receive a copy from the school where they applied. 7 The instruction for making the preference list, reads as follows (translated from the Dutch instruction): We would like to know your preferences for schools in case no single school would conduct a lottery. You should, therefore, not consider the possibility of losing a lottery. This may imply that you place another school at place 1 than the school where you applied. You can only list schools that offer the level of education that corresponds with the recommended level of your child. would qualify. For each school we ask you to award points. Hence schools for which your child The highest-ranked school receives 100 points. The points you give to another school, can be seen as a percentage of the highest-ranked school. The lower a school is on your list, the fewer points you award. If a school on the list is very close to the previous one, the difference in points is small. For clarity we give some examples. The examples make clear that the difference in points that are awarded to schools should reflect the difference in valuation that is given to schools. By setting the maximum score for all students to 100 and the minimum to zero we make sure all students have the same range. Appendix A contains a translation of the examples from Dutch. Figure 1 shows a preference list submitted by one of the respondents. maximum number of schools that students could include in their preference list equals ten. Note that this restriction does not interfere with truth-telling as would be the case if a preference list submitted for the DA mechanism is restricted in length (e.g. Calsamiglia et al., 2010). We clearly stated that we would like to know preferences for schools in case no single school would conduct a lottery. We can, therefore, assume that the preference list of the questionnaire coincides with the (at most) ten highest-ranked schools that would be submitted under a DA system without restrictions on the length of the preference list. 7 This applies, for example, to students who were enrolled in a primary school outside Amsterdam and applied to a secondary school in Amsterdam. The 10

Figure 1. Example of preference list Note: The translation from Dutch says: Preferences for schools (in accordance with the advised track level). Response The response rate to the questionnaire was 47 percent. This is far above the 20 percent response rate that the Research and Statistics unit of the municipality achieved on a questionnaire about school choice (see Cohen et al., 2012). Our response rate varies substantially between students from different track levels and is highest for students with an advice for the pre-university track. The response rate for this group is 64 percent. Because the assignment and possible lotteries are done separately for students from different track levels, we focus our analysis on this group with the highest response rate. In 2013, 1923 students chose among 29 schools in Amsterdam that offer the pre-university track. Four schools were oversubscribed in the first round and 80 students lost a lottery. Figure 2 shows the location of the 29 schools and the location of the home addresses of students that applied to these schools. As can been seen in Figure 2, many schools are located close to each other in the center of Amsterdam while the students come from all over Amsterdam and some even from outside Amsterdam. Table 1 reports descriptive statistics for all students with an advice at the preuniversity level in 2013, separately for students who responded and students who did not respond to the questionnaire. This shows that girls are more likely to respond than boys (p=0.051). Response is also higher among students with a higher score on the final exit test from primary school (p=0.000). Students who will lose a lottery (i.e. apply to a popular school) are more likely to respond than students who will 11

Figure 2. Location of schools and students in the academic track in Amsterdam Note: The blue dots are home addresses of students, the orange dots are the location of schools. 12

Table 1. Summary statistics Responded (1240) Did not respond (683) p-value Mean SD Nobs Mean SD Nobs Girl (dummy) 0.49 0.50 1236 0.44 0.50 673 0.051 Test score (scale 500-550) 546.68 3.23 1196 546.07 3.29 622 0.000 Student lost the lottery (dummy) 0.05 0.22 1240 0.02 0.16 683 0.002 Student background: Neighborhood income in top 75% 0.70 0.46 1240 0.73 0.44 683 0.122 Neighborhood income in bottom 25% 0.26 0.44 1240 0.21 0.41 683 0.015 Neighborhood income unknown 0.04 0.20 1240 0.06 0.23 683 0.138 Notes: The table presents means, standard deviations and number of observations of selected variables for students with a pre-university advice, by response status. Test score is the score on the nationwide exit test from primary school. The score runs from 500 to 550. Students admitted to the pre-university track usually have a score exceeding 542. Standard deviation of test score in the entire sample equals 11.25. Neighborhood income is monthly average taxable income in the (6-digit) postal code area of residence of the student, measured in December 2008. The top 75% and bottom 25% are based on the sample of students in the pre-university track in Amsterdam in 2013. The p-values in the final column are for tests of the equality of the variable for respondents and non-respondents. not lose a lottery (p=0.002). Table 2 lists the 29 schools that offer the pre-university track in Amsterdam and provides relevant information. Columns (1) to (3) report information about the number of students that applied to that school in the first round, the capacity of the school and the number of applicants that responded to our questionnaire. 8 The response rate to the questionnaire varies between schools, which can partly be explained by how persistent the local administrator asked for the completed questionnaire at the moment the student applied for this school. Column (4) shows the percentage of the respondents to our questionnaire, without priority at the school of application, who expect a lottery at their school of application, while columns (5) and (6) indicate whether the school actually conducted a lottery among its applicants at the pre-university level the year of our study and the previous year. Students have difficulties in predicting whether a school will be oversubscribed or not. For example, 86.6 percent of the respondents who applied to Vossius predict that there is a risk that they will loose the lottery at this school, while Vossius was not oversubscribed this year (24 of the 150 seats were vacant after round 1), nor in the previous year. 9 8 Taking the sum over the positive differences between registrations and capacity gives a total of 92 students. This number is larger than the number of students that lost a lottery (80). This is due to the fact that some students who applied for a school changed their mind after the end of the application period but before the lotteries were conducted, or were not allowed to enroll in the school to which they applied because their test score was too low. 9 Papers that use a structural approach to estimate the degree of strategic behavior assume that 13

Figure 3. Distribution of number of schools on preference lists.3 Share of students.2.1 0 0 1 2 3 4 5 6 7 8 9 10 Number of schools on preference list Note: Figure is based on 1240 students with an advice for the pre-university track who responded to the questionnaire. The final two columns report the GPA and pass rate per school of the cohort graduating in 2012; GPA runs on a scale from one to ten. We took this information from the website of the school inspectorate, which is publicly accessible. These are uncorrected indicators of school quality, which students may use when they choose schools. There are clear differences in GPA and pass rate between schools. Schools with a GPA above 6.5 are schools that exclusively offer the pre-university track. Respondents could include up to ten schools in their preference list. Figure 3 shows the distribution of the numbers of schools that were actually included in the submitted preference lists. Up to 20 percent of the respondents mention at most two schools. Almost 30 percent of the respondents rank three schools and the remaining 50 percent list at least four schools. Few students fill the complete list of ten schools. In the subsequent analyses we need to deal with non-response to the questionnaire and with incomplete preference lists. When we analyze strategic behavior, we use sampling weights. To construct these weights, we estimate a logit model of response status using the test score, gender and school of application in the first round as predictors. The sampling weights are equal to the inverse of the predicted probability of response. By using these sampling weights, our estimates of strategic behavior pertain to the population of students at the pre-university level applying to a secondary school in Amsterdam, and not only to those who responded to the questionnaire. The key advantage of using a logit specification with school fixed all or a share of the students have correct beliefs about the choices of others. Table 2 suggests that in our setting this is a strong assumption as students have difficulties predicting which schools will be oversubscribed. 14

Table 2. The schools Name school Registrations Capacity Responders % predicts lottery c Lottery this year Lottery last year GPA Pass rate (1) (2) (3) (4) (5) (6) (7) (8) HAL 203 171 122 83.5 Y 6.4 0.9 Hyperion a 177 189 143 70.5 na na Ignatius 171 145 158 95.5 Y Y 6.8 0.92 Cygnus 160 175 89 54.7 6.8 0.91 Barlaeus 151 140 87 87.8 Y 6.8 0.91 4e gymnasium b 143 140 105 64.5 6.6 0.87 Vossius 126 150 80 86.6 6.8 0.95 St.Nicolaas 118 95 112 62.6 Y 6 0.81 Caland 77 84 42 27.0 6.1 0.89 Fons Vitae 70 69 30 70.8 6 0.87 Spinoza 68 78 45 88.9 Y 6.3 0.94 Damstede 60 59 26 33.3 6.2 0.85 MLA 56 112 38 na 6 0.92 Berlage 47 112 21 42.9 6.2 0.77 Nieuwland 46 56 21 27.3 5.6 0.65 Geert Groote 43 60 14 na 6.2 0.94 HL Zuid 34 50 12 na 6 0.95 Gerrit vd Veen 31 48 13 na Y 5.7 0.77 Cartesius 29 70 20 57.1 6.2 0.95 Reigersbos 23 25 17 23.5 5.9 1 IJburg a 16 50 7 na na na Outside Amsterdam 15 1000 6 na Comenius 13 50 10 na 6 0.84 OSB 10 100 5 na 5.4 0.77 Cosmicus 9 48 2 na 4.1 0.73 HL West 8 50 5 na 5.6 0.69 CSB 5 0 5 na 5.6 0.75 Bredero 3 10 2 na 5.7 0.75 Maimonides 2 20 2 na 5.8 0.84 Note: a Information about pass rates and GPA were not available in 2012 for Hyperion and IJburg, because these are new schools and no student in the academic track graduated from these schools before 2013. b For some schools the number of registrations exceeds capacity, but there is no lottery. In these cases some students who applied for the school changed their mind just after round one and applied for another school or they were not allowed to enroll in the school, because their test score was too low. c This column shows the percentage of the respondents, without priority at the school of application, that predicts a lottery at their school of application. It is only reported for schools that have more than 10 respondents without priority that answered question 4 of the survey. 15

effects is that for each school the weighted sum of survey respondents will exactly match the total number of applicants. When simulating the different assignment models, we need information on the full population of students. Therefore, for each student who did not complete the survey we sample a replacement student from the pool of students who responded to the survey and applied to the same school. The sampling of a specific student is done using the predicted probabilities from the logit model. 10 Next, there are quite some students who submitted relatively few schools on their preference list. When students submitted lists with less than ten schools, we complete their lists by sampling additional schools from the distribution of preferences of students applying to the same school. The probability that a specific school is added to the preference list is proportional to how often this school occurs on preference lists of students applying for the same school. For each student we add schools until the preference list contains ten unique schools. The newly sampled schools are ranked but the number of preference points is set equal to zero. This imputation procedure will result in conservative estimates of the performance of the DA mechanism. A student who is assigned to a newly sampled school can occupy the place of another student who ranks that school high and who assigns positive points to that school. Appendix B shows results from the logit model that was used to create sampling weights and propensity scores. In the Amsterdam version of the Boston mechanism students apply to one school and do not submit a ranked list of other schools. Such lists are, however, needed to simulate the Boston mechanism. For each student we completed their Boston list using information from their preference list. Practically, if a student submitted a preference list with (in order) schools s 1, s 2, s 3,..., s 10, and the student applied to school s 2, then we assign this student the following ordered Boston list: s 2, s 1, s 3,..., s 10. This is not necessarily the list that this student would submit if she were to submit a Boston list. When she strategically does not apply to her most-preferred school s 1, she might not rank this school second on the Boston list. Our procedure to complete the Boston list, helps Boston to perform well. As we will be shown below, Boston places almost 95 percent of the students at the school where they applied. Students who do not receive a place at the school where they applied, do with our procedure still make a chance at their most-preferred schools if it happens 10 For all survey respondents applying to the same school we take the difference in predicted survey-responding probability with the student that should be replaced. Next, we randomly draw one student as replacement, where the replacement probabilities are proportional to a Gaussian kernel in the difference in predicted survey-responding probabilities with a bandwidth of 0.05. The random draws are with replacement and all results are robust with respect to the bandwidth choice. 16

that these school were not oversubscribed in the first round. The data imputation procedure is performed 60 times. After each data imputation the different school mechanisms are simulated 60 times. Our results are, therefore, based on a total of 3600 simulations. Standard errors are based on variation of outcomes when simulating the mechanisms, but not on variation of outcome due to data imputation. The reason is that the focus of this paper is on comparing school assignment mechanisms. 5 Results The results are presented in four subsections. Subsection 5.1 reports results on strategic behavior. The second subsection compares the performance of Boston to DA-STB and DA-MTB on the basis of ex-post efficiency, while Subsection 5.3 compares the mechanisms on the basis of ex-ante efficiency. The final subsection, 5.4, reports how outcomes from the different mechanisms are related to students social background. 5.1 Strategic behavior Table 3 reports the share of students who did not apply to the school which they ranked first on their preference list. 11 The shares in the first row are obtained by weighting with the inverse of the estimated probability to respond to the questionnaire. The shares in the second row are unweighted, and these are slightly lower than the weighted shares. The share of students disguising their true preferences equals eight percent (column (1)), and is slightly (but not significantly; p=0.763) higher for boys than for girls (columns (2) and (3)). There is also no significant difference (p=0.762) in strategic behavior between students from different social backgrounds (columns (4) and (5)). We define social background on the basis of average neighborhood income, with disadvantaged (non-disadvantaged) students living in a street where the average taxable income is in the bottom 25 (top 75) percent of the distribution. 12 Columns (7) and (8) shows that students whose most-preferred school turns out to be oversubscribed, are more likely to apply to another school than their most-preferred school. This indicates that oversubscription is partly predictable, but recall from Table 2 that it is certainly not fully predictable. 11 Students who choose strategically place the school where they apply, on average, on place 2.32 on their preference list. 12 Average neighborhood income at the 6-digit postal code area is reported by Statistics Netherlands in December 2008. A 6-digit postal code covers a street or part of a street. The top 75% and bottom 25% are based on the sample of students with an advice for the pre-university track in Amsterdam in 2013. 17

Table 3. Share of students that did not apply for their most preferred school All Gender Social background Lottery at most preferred Difference points 1st & 2nd school a Boys Girls Non-disadv Disadv Unknown Yes No 0-5 6-10 11-20 21-100 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) weighted 0.078 0.081 0.075 0.078 0.084 0.038 0.095 0.068 0.122 0.057 0.073 0.052 share (0.009) (0.013) (0.013) (0.011) (0.019) (0.026) (0.015) (0.011) (0.022) (0.016) (0.020) (0.015) unweighted 0.065 0.066 0.065 0.066 0.068 0.038 0.077 0.057 0.114 0.045 0.056 0.041 share (0.007) (0.010) (0.010) (0.008) (0.014) (0.026) (0.012) (0.009) (0.019) (0.012) (0.014) (0.011) N 1214 622 588 852 309 53 495 719 271 292 288 318 Note: Weighted shares are weighted by the inverse of the predicted probability of response. Robust standard errors between parentheses. All estimates are significant at the 1%-level except those in column (6). a Difference in preference points is set to 100 if a second school is listed on the preference list but no points were awarded to this second school. 18

Theory predicts that students with strong preferences for their most-preferred school are less likely to behave strategically than students with weaker preferences (Abdulkadiroǧlu et al., 2011). To inquire this, we divide the sample into four (almost) equally sized groups, based on the difference in preference points between the first and second ranked school on students preference lists. For the first group this difference is between zero and five points and then increases to six to ten (second group), to 11 to 20 (third group), and to 21 to 100 for the fourth group. Columns (9) to (12) of Table 3 report the shares of students that did not apply to their most-preferred school for each of the four groups. In accordance with the theoretical prediction, the share of students not applying to their most-preferred school is largest in the first group (12.2 percent) and smallest in the fourth group (5.2 percent). The ranking for the second and third groups is not monotonic, but the difference in share between these groups is small and not statistically significant (p=0.52). 13 We, therefore, conclude, that the intensity of preferences indeed plays a role for choices made under the Boston mechanism. Students who have a strong preference for their most-preferred school are less likely to choose strategically than students who are almost indifferent between their first and second ranked schools. This is an important finding, as it fulfills the requirement for the Boston mechanism to have the potential to outperform DA in terms of welfare (ex-ante efficiency). 5.2 Ex-post efficiency: Boston versus DA-STB and DA-MTB In this subsection we compare the performance of Boston with that of DA-STB and DA-MTB on the basis of ex-post efficiency. We compare the mechanisms on the following dimensions: i) shares of students placed in one their most-preferred n schools; ii) shares of students that are ex-post better, equal or worse off under Boston than under DA-STB and DA-MTB; iii) Shares of students that would like to switch places without harming other students; iv) Shares of students that may regret their choices made under Boston. The results in this subsection are only based on students ordinal preferences regarding schools. Placement in most-preferred schools. Table 4 compares at which school on their preference list students are placed under Boston, DA-STB and DA-MTB. The first row shows that Boston performs best in terms of placement of students at the school of application. This is not surprising as Boston explicitly maximizes this measure. The results in the second row show that placement at school of application is a 13 The difference between the first and last group is significant (p=0.009). 19

Table 4. Placement and mean preference points Boston DA-STB DA-MTB Application 0.947 (0.000) 0.828 (0.004) 0.780 (0.008) 1st choice 0.861 (0.002) 0.892 (0.003) 0.833 (0.008) 2nd choice 0.073 (0.003) 0.064 (0.004) 0.138 (0.007) 3rd choice 0.030 (0.002) 0.024 (0.003) 0.024 (0.004) 4th choice 0.016 (0.002) 0.011 (0.002) 0.004 (0.002) 5th choice 0.008 (0.001) 0.005 (0.002) 0.001 (0.001) 6th choice 0.004 (0.001) 0.002 (0.001) 0.000 (0.000) 7th choice 0.002 (0.000) 0.001 (0.001) 0.000 (0.000) 8th choice 0.002 (0.000) 0.000 (0.000) 0.000 (0.000) 9th choice 0.002 (0.000) 0.000 (0.000) 0.000 (0.000) 10th choice 0.001 (0.000) 0.000 (0.000) 0.000 (0.000) Mean points 92.72 (0.180) 93.81 (0.243) 93.42 (0.312) Prob. larger than Boston 0.998 (0.015) 0.934 (0.171) Notes: The table reports results of 3600 simulations. Numbers in parentheses are standard errors misleading performance measure. Because a substantial share of students do not apply to their most-preferred school under Boston, the share of students placed in their most-preferred school is (significantly) lower under Boston than under DA- STB. Also the difference of 16 percentage points between Boston and DA-MTB in the first row shrinks to three percentage points in the second row. DA-MTB performs worse than Boston and DA-STB in placing students in their most-preferred school. It does, however, better for the second most preferred school. Figure 4 shows the results of Table 4 cumulatively. For example, the solid line indicates that under Boston 86 percent of the students is placed in their mostpreferred school, 93 percent in their first or second most-preferred school, 96% in their first, second or third most preferred school, and so on. Two things in Figure 4 stand out. The first is that DA-STB stochastically dominates Boston; at every value for rank on the preference list, the share that is placed under DA-STB exceeds that of Boston. The second is that there is a trade off between DA-MTB and DA-STB (and Boston). While DA-STB assigns more students to their most-preferred school than DA-MTB, DA-MTB places a larger fraction of students in one of their two (and higher) most-preferred schools. Abdulkadiroǧlu et al. (2009) document a similar pattern for New York City where DA-STB does better than DA-MTB for the seven (and lower) most-preferred schools, and DA-MTB does better than DA-STB for the eight (and higher) most-preferred schools. The reason for this pattern with a single crossing point is as follows. Under 20

Figure 4. Cumulative distribution of students to schools on their preference list 1.98 Cumulative share of students.96.94.92.9.88.86.84.82.8 Boston DA with multiple tie-breaking DA with single tie-breaking 1 2 3 4 5 6 7 8 9 10 Rank on school preference list both DA-STB and DA-MTB students who lose the lottery at their first-ranked school are the ones who drew an unfavorable lottery number. When DA-STB is in place, first-round losers carry their bad draw with them to their second-ranked school. When DA-MTB is in place, losers in the first round draw a new and in expectation more favorable lottery number at their second-ranked school, making it more likely with DA-MTB than with DA-STB that they capture the place from a student who was tentatively assigned to her first-ranked school in the first round. See Arnosti (2015) for a theoretical discussion of this result. Shares of winners and losers. The next dimension on which we compare the performance of Boston vis-a-vis the two versions of DA are the shares of students that are better, equal or worse off with Boston than with DA in terms of ex-post assignment. We register after each simulation whether a student gets a seat in a higher, equal or worse-ranked school under Boston than under DA, and compute the respective shares across the 3600 simulations. Our measure for ex-post efficiency is thus based on an ordinal measure for preferences. The average of these shares over all students are reported in columns (1) and (2) of Table 5. The ex-post efficiency results in Table 5 show that, on average, students have a 12 percent and 11 percent probability to get a seat in a higher-ranked school 21